examples/l3fwd/l3fwd_lpm_neon.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 Intel Corporation.
   3  * Copyright(c) 2017-2018 Linaro Limited.
   4  */
   5
   6 #ifndef __L3FWD_LPM_NEON_H__
   7 #define __L3FWD_LPM_NEON_H__
   8
   9 #include <arm_neon.h>
  10
  11 #include "l3fwd_neon.h"
  12
  13 /*
  14  * Read packet_type and destination IPV4 addresses from 4 mbufs.
  15  */
  16 static inline void
  17 processx4_step1(struct rte_mbuf *pkt[FWDSTEP],
  18                 int32x4_t *dip,
  19                 uint32_t *ipv4_flag)
  20 {
  21         struct rte_ipv4_hdr *ipv4_hdr;
  22         struct rte_ether_hdr *eth_hdr;
  23         int32_t dst[FWDSTEP];
  24
  25         eth_hdr = rte_pktmbuf_mtod(pkt[0], struct rte_ether_hdr *);
  26         ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
  27         dst[0] = ipv4_hdr->dst_addr;
  28         ipv4_flag[0] = pkt[0]->packet_type & RTE_PTYPE_L3_IPV4;
  29
  30         eth_hdr = rte_pktmbuf_mtod(pkt[1], struct rte_ether_hdr *);
  31         ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
  32         dst[1] = ipv4_hdr->dst_addr;
  33         ipv4_flag[0] &= pkt[1]->packet_type;
  34
  35         eth_hdr = rte_pktmbuf_mtod(pkt[2], struct rte_ether_hdr *);
  36         ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
  37         dst[2] = ipv4_hdr->dst_addr;
  38         ipv4_flag[0] &= pkt[2]->packet_type;
  39
  40         eth_hdr = rte_pktmbuf_mtod(pkt[3], struct rte_ether_hdr *);
  41         ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
  42         dst[3] = ipv4_hdr->dst_addr;
  43         ipv4_flag[0] &= pkt[3]->packet_type;
  44
  45         dip[0] = vld1q_s32(dst);
  46 }
  47
  48 /*
  49  * Lookup into LPM for destination port.
  50  * If lookup fails, use incoming port (portid) as destination port.
  51  */
  52 static inline void
  53 processx4_step2(const struct lcore_conf *qconf,
  54                 int32x4_t dip,
  55                 uint32_t ipv4_flag,
  56                 uint16_t portid,
  57                 struct rte_mbuf *pkt[FWDSTEP],
  58                 uint16_t dprt[FWDSTEP])
  59 {
  60         rte_xmm_t dst;
  61
  62         dip = vreinterpretq_s32_u8(vrev32q_u8(vreinterpretq_u8_s32(dip)));
  63
  64         /* if all 4 packets are IPV4. */
  65         if (likely(ipv4_flag)) {
  66                 rte_lpm_lookupx4(qconf->ipv4_lookup_struct, dip, dst.u32,
  67                         portid);
  68                 /* get rid of unused upper 16 bit for each dport. */
  69                 vst1_s16((int16_t *)dprt, vqmovn_s32(dst.x));
  70         } else {
  71                 dst.x = dip;
  72                 dprt[0] = lpm_get_dst_port_with_ipv4(qconf, pkt[0],
  73                                                      dst.u32[0], portid);
  74                 dprt[1] = lpm_get_dst_port_with_ipv4(qconf, pkt[1],
  75                                                      dst.u32[1], portid);
  76                 dprt[2] = lpm_get_dst_port_with_ipv4(qconf, pkt[2],
  77                                                      dst.u32[2], portid);
  78                 dprt[3] = lpm_get_dst_port_with_ipv4(qconf, pkt[3],
  79                                                      dst.u32[3], portid);
  80         }
  81 }
  82
  83 /*
  84  * Buffer optimized handling of packets, invoked
  85  * from main_loop.
  86  */
  87 static inline void
  88 l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
  89                         uint16_t portid, struct lcore_conf *qconf)
  90 {
  91         int32_t i = 0, j = 0;
  92         uint16_t dst_port[MAX_PKT_BURST];
  93         int32x4_t dip;
  94         uint32_t ipv4_flag;
  95         const int32_t k = RTE_ALIGN_FLOOR(nb_rx, FWDSTEP);
  96         const int32_t m = nb_rx % FWDSTEP;
  97
  98         if (k) {
  99                 for (i = 0; i < FWDSTEP; i++) {
 100                         rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
 101                                                         void *));
 102                 }
 103
 104                 for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
 105                         for (i = 0; i < FWDSTEP; i++) {
 106                                 rte_prefetch0(rte_pktmbuf_mtod(
 107                                                 pkts_burst[j + i + FWDSTEP],
 108                                                 void *));
 109                         }
 110
 111                         processx4_step1(&pkts_burst[j], &dip, &ipv4_flag);
 112                         processx4_step2(qconf, dip, ipv4_flag, portid,
 113                                         &pkts_burst[j], &dst_port[j]);
 114                 }
 115
 116                 processx4_step1(&pkts_burst[j], &dip, &ipv4_flag);
 117                 processx4_step2(qconf, dip, ipv4_flag, portid, &pkts_burst[j],
 118                                 &dst_port[j]);
 119
 120                 j += FWDSTEP;
 121         }
 122
 123         if (m) {
 124                 /* Prefetch last up to 3 packets one by one */
 125                 switch (m) {
 126                 case 3:
 127                         rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
 128                                                         void *));
 129                         j++;
 130                         /* fallthrough */
 131                 case 2:
 132                         rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
 133                                                         void *));
 134                         j++;
 135                         /* fallthrough */
 136                 case 1:
 137                         rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
 138                                                         void *));
 139                         j++;
 140                 }
 141
 142                 j -= m;
 143                 /* Classify last up to 3 packets one by one */
 144                 switch (m) {
 145                 case 3:
 146                         dst_port[j] = lpm_get_dst_port(qconf, pkts_burst[j],
 147                                                        portid);
 148                         j++;
 149                         /* fallthrough */
 150                 case 2:
 151                         dst_port[j] = lpm_get_dst_port(qconf, pkts_burst[j],
 152                                                        portid);
 153                         j++;
 154                         /* fallthrough */
 155                 case 1:
 156                         dst_port[j] = lpm_get_dst_port(qconf, pkts_burst[j],
 157                                                        portid);
 158                 }
 159         }
 160
 161         send_packets_multi(qconf, pkts_burst, dst_port, nb_rx);
 162 }
 163
 164 #endif /* __L3FWD_LPM_NEON_H__ */